شباهت ها و تفاوت بین ادراکات از بزهکاری همسالان، بزهکاری خود گزارشی همسالان و بزهکاری مخاطب: تجزیه و تحلیل از زوج دوستی

Abstract Purpose Despite the peer delinquency measurement debate having profound implications for research, looming questions remain about the validity of various forms of peer delinquency operationalizations. This study examines whether perceptions of peer delinquency form identical latent constructs to both respondent and peer self-reported delinquency. Methods Using a dyadic dataset of friendship pairs, confirmatory factor analysis and model comparison tests are used to determine the degree of similarity between perceptions of peer delinquency, respondent self-reported delinquency, and peer self-reported delinquency. Results Peer self-reports and respondent perceptions of peer delinquency load on unique constructs across a number of different behaviors. For most behaviors, respondent perceptions of peer delinquency and respondent self-reports load on separate constructs. Results further indicate that respondent and peer self-reported delinquency are separate latent constructs. Finally, the strength of the association between respondent and peer delinquency is substantively smaller in magnitude, by as much as a factor of three in some instances, when peer delinquency is measured via peer self-reports in place of respondent perceptions. Conclusions Reports of peer delinquency provided directly from peers demonstrate strong discriminant validity in relation to self-reported delinquency, while perceptions of peer delinquency demonstrate poorer discriminant validity, particularly for theft and violence constructs.

Introduction The correlation between peer delinquency and respondent delinquency is one of the most replicated findings within criminology (Akers, 2009 and Pratt et al., 2010). Research consistently demonstrates that the strength of this relationship meets or exceeds the influence of other theoretical variables, including self-control and variables from the rational choice/deterrence traditions (Pratt et al., 2010). In fact, researchers have noted that “Next to prior delinquency, association with delinquent peers is the best predictor of delinquency” (Agnew, 1991a, p. 148). Despite this, a critical debate centers on the measurement of the peer delinquency construct (e.g., Rebellon and Modecki, in press and Young et al., 2011). Although Akers (2009) has argued that the most appropriate form of peer delinquency measurement should be perceptually-based, many have argued against the use of such measures (e.g., Haynie & Osgood, 2005) and instead have favored the use of direct reports of peer delinquency to avoid issues such as projection, which is the situation when a respondent reports his/her own delinquency when perceiving how delinquent his/her peers have behaved (see Boman et al., 2012 and Young et al., 2011). Drawing on the projection concept, some have argued that there are minimal differences between the constructs of perceptual peer delinquency and self-reported delinquency (Gottfredson & Hirschi, 1990). If such a claim is true, then “virtually the entire body of research assessing the relationship between association with deviant peers and self-reported deviance would be useless because the measures are not independent” (Thornberry & Krohn, 1997, p. 222). A small body of research has examined the degree of similarity between perceptions of peer delinquency and self-reported delinquency using latent variable approaches. For example, using data from multiple waves of the National Youth Survey (NYS), Rebellon (2012) found that there were statistical (though not substantive) differences between self-reported substance use indicators and perceptions of peers’ substance use. Similar conclusions have been reached when considering property offenses (Zhang & Messner, 2000). In short, there appears to be evidence that partially supports the critique that perceptions of peer delinquency lack strong discriminant validity when examined in relation to self-reports of delinquency. More recently, researchers have started to consider the discriminant validity between perceptual measures of peer delinquency and reports of peer delinquency gathered directly from the peer him/herself, which are becoming a frequently used alternative to the traditional “perceptual” measurement strategy (these are often called “direct” reports of peer delinquency; Rebellon and Modecki, in press and Young et al., in press). This is an important methodological consideration, as the first logical step when developing any particular scale is to evaluate if the scale is measuring what it claims to be measuring. Applying this intuition to the peer delinquency construct, the most important step is to evaluate if perceptions are measuring actual peer delinquency. The fact that researchers have only recently started to direct attention to this issue in part reflects the difficulty of obtaining measures of peer self-reported delinquency – they require the use of dyadic or social-networking methodologies (Boman et al., 2012, Meldrum et al., 2009 and Rebellon and Modecki, in press). However, with the increasing availability of such data, determining whether perceptions of peer delinquency are reflective of actual peer behavior is rapidly becoming realized as a critical area of research. Despite the importance of recent work in these areas of research, latent variable investigations into discriminant validity between self-reported delinquency and each of the two measures of peer delinquency (perceptions and direct reports) have been limited to only a handful of empirical studies (Rebellon, 2012, Rebellon and Modecki, in press, Young et al., in press and Zhang and Messner, 2000), each with shortcomings that warrant additional research. For example, some studies have been limited to a focus on perceptual measures of peer delinquency in relation to self-reported delinquency without consideration given to direct reports (Rebellon, 2012 and Zhang and Messner, 2000). In addition, of the studies which have included both direct and perceptual measures of peer delinquency in the same analyses, they have been limited with regard to the scope of the behaviors considered (Rebellon and Modecki, in press and Young et al., in press). Discriminant validity may vary according to the type of behavior investigated, and there are a number of additional constructs that should be considered in order to assess the generality of findings stemming from previous research. As Warr (2002) points out, certain types of delinquency – namely substance use behaviors – are more likely to be shared in common between friends. On the other hand, acts of theft and violence are often committed without friends. It is possible, then, that the discriminant validity between perceptions of peer delinquency, direct reports of peer delinquency, and self-reported delinquency could depend upon the specific behavior under consideration. In an effort to advance this area of research, build upon recent work, and address the aforementioned considerations, the current study makes use of dyadic data collected from a large sample of young adults in friendship pairs at a university located in the southeastern United States. The data contain information on perceptual peer delinquency, self-reported respondent delinquency, and direct reports of peer delinquency for nearly two dozen behaviors. Thus, the data are well-suited for contributing to the existing literature and advancing our understanding of similarities and differences between these three measures of delinquency.

Results Comparing actor perceptions of peer delinquency to partner self-reported delinquency Table 2 reports the first series of CFA models which compare one- and two-factor solutions of the actor’s perception of peer delinquency and the partner’s own self-reported delinquency. All factor loadings are standardized. The first latent construct under consideration is theft, and the construct’s results are presented in the far left set of columns. This first set of models provides initial support for hypothesis one in that there is a clear distinction between actor perceptions of peer theft and the peer’s own self-reported theft. The fit of the two-factor model outperforms the one-factor model (χ2 = 354.319 in the one-factor model; χ2 = 49.346 in the two-factor model), and the two-factor correlation is well below 1.00 (r = 0.496). Table 2. Confirmatory Factor Analysis Models Loading Actor Perceptions of Peer Delinquency (APPD) and Peer Self-Reported Delinquency (PSRD) Measures onto One and Two Constructs (n = 2,154) Theft Vandalism Violence Drugs Alcohol 1 Factor 2 Factor 1 Factor 2 Factor 1 Factor 2 Factor 1 Factor 2 Factor 1 Factor 2 Factor Item # λ λ λ λ λ λ λ λ λ λ APPD 1 .822 .899 / - .885 .926 / - .713 .766 / - .823 .897 / - .852 .906 / - APPD 2 .806 .851 / - .883 .923 / - .746 .815 / - .798 .822 / - .715 .741 / - APPD 3 .703 .748 / - .914 .939 / - .629 .704 / - .821 .850 / - .637 .663 / - APPD 4 .568 .645 / - .706 .787 / - .852 .873 / - .829 .857 / - APPD 5 .614 .695 / - .884 .886 / - PSRD 1 .704 - / .841 .672 - / .784 .615 - / .680 .889 - / .954 .851 - / .892 PSRD 2 .781 - / .864 .806 - / .885 .658 - / .739 .743 - / .757 .668 - / .694 PSRD 3 .437 - / .549 .803 - / .902 .610 - / .696 .818 - / .844 .609 - / .633 PSRD 4 .494 - / .575 .490 - / .556 .846 - / .848 .851 - / .876 PSRD 5 .446 - / .538 .729 - / .754 2 Factor Correlation .496⁎⁎⁎ .398⁎⁎⁎ .611⁎⁎⁎ .829⁎⁎⁎ .838⁎⁎⁎ 1 Factor Model Statistics χ2 354.319⁎⁎⁎ 606.742⁎⁎⁎ 248.152⁎⁎⁎ 240.847⁎⁎⁎ 586.157⁎⁎⁎ CFI .813 .871 .858 .942 .954 TLI .813 .856 .840 .957 .969 RMSEA .110 .131 .122 .083 .137 2 Factor Model Statistics χ2 49.346⁎⁎⁎ 49.940⁎⁎⁎ 39.234⁎⁎⁎ 205.684⁎⁎⁎ 226.867⁎⁎⁎ CFI .981 .993 .981 .952 .983 TLI .981 .993 .975 .968 .989 RMSEA .035 .028 .048 .072 .084 Model Comparison Test χ2 143.786⁎⁎⁎ 210.744⁎⁎⁎ 118.543⁎⁎⁎ 45.771⁎⁎⁎ 236.582⁎⁎⁎ Note: All factor loadings significant at p ≤ .001. ⁎⁎⁎ p ≤ .001. Table options For vandalism, the fit indices of the two-factor model show evidence of close fit to the data, and the two-factor model outperforms the one-factor model based on the model comparison test (Δ χ2 = 210.744, p < .001). As in the theft models, the latent constructs of actor perceived vandalism and peer self-reported vandalism are moderately correlated (r = .398). Analogous results are found in the comparison of the violence models; the model comparison χ2 value is large (χ2 = 118.543) and indicates the two-factor model provides a closer fit to the data than the one-factor model. On the other hand, the correlation between the latent constructs of perceptual and peer self-reported violence is stronger than for the prior two constructs (r = .611). The final two comparisons in Table 2 provide slightly different results from the previous models. For the latent construct of drug-related behaviors, there is a strong similarity between the two factors of actor perceptions and peer self-reports (r = .829), but the two-factor model fits the data more closely than the one-factor model (Δ χ2 = 45.771). For the alcohol-related behavior construct, the perceptual and peer self-reported constructs are again very similar (r = .838) even though the model comparison test for the alcohol model shows that the two-factor solution provides a substantively closer fit to the data than the one-factor model (Δ χ2 = 236.582). Taken together, the models in Table 2 provide support for hypothesis one and indicate that actor perceptions of peer delinquency are distinct constructs from the peer’s self-reported delinquency. That being said, actor perceptions of the peer’s substance use-related behaviors do reflect the peer’s self-reported behavior with a rather high level of precision. Comparing actor perceptions of peer delinquency to actor self-reported delinquency Table 3 reports the second series of CFA models which compare one- and two-factor solutions of the actor’s perception of peer delinquency and the actor’s own self-reported delinquency. Starting with the theft construct, the item loadings for the perceptions and self-reports of theft in both the one-factor and two-factor models are very similar. Further, the model fit indices between the two constructs are also extremely similar and, in fact, are so similar that the model comparison test reveals that the two-factor model does not significantly outperform the one-factor model (Δ χ2 = 1.228, p = .27). Thus, the latent constructs of perceptual peer theft and actor self-reported theft in these data are indistinguishable from one another. In addition to loading on the same latent construct, the correlation between actor perceptions and self-reported theft behavior approach 1.00 in the two-factor solution (r = .973). This suggests that perceptions of peer theft are not substantively different from the actor’s own reported theft. Table 3. Confirmatory Factor Analysis Models Loading Actor Perceptions of Peer Delinquency (APPD) and Actor Self-Reported Delinquency (ASRD) Measures onto One and Two Constructs (n = 2,154) Theft Vandalism Violence Drugs Alcohol 1 Factor 2 Factor 1 Factor 2 Factor 1 Factor 2 Factor 1 Factor 2 Factor 1 Factor 2 Factor Item # λ λ λ λ λ λ λ λ λ λ APPD 1 .804 .809 / - .885 .902 / - .752 .755 / - .747 .788 / - .824 .864 / - APPD 2 .800 .803 / - .911 .929 / - .755 .757 / - .846 .871 / - .727 .753 / - APPD 3 .766 .769 / - .917 .930 / - .773 .776 / - .841 .866 / - .676 .702 / - APPD 4 .793 .800 / - .805 .835 / - .873 .893 / - .844 .882 / - APPD 5 .710 .736 / - .902 .913 / - ASRD 1 .733 - / .740 .750 - / .789 .676 - / .678 .843 - / .873 .842 - / .862 ASRD 2 .748 - / .754 .811 - / .846 .672 - / .674 .766 - / .787 .674 - / .698 ASRD 3 .622 - / .627 .818 - / .853 .757 - / .761 .825 - / .851 .666 - / .690 ASRD 4 .745 - / .752 .678 - / .713 .879 - / .893 .880 - / .910 ASRD 5 .592 - / .619 .804 - / .829 2 Factor Correlation .973⁎⁎⁎ .844⁎⁎⁎ .987⁎⁎⁎ .877⁎⁎⁎ .856⁎⁎⁎ 1 Factor Model Statistics χ2 306.666⁎⁎⁎ 408.967⁎⁎⁎ 304.462⁎⁎⁎ 194.316⁎⁎⁎ 887.434⁎⁎⁎ CFI .915 .894 .886 .961 .919 TLI .927 .939 .886 .974 .954 RMSEA .098 .101 .145 .064 .176 2 Factor Model Statistics χ2 310.055⁎⁎⁎ 376.148⁎⁎⁎ 303.728⁎⁎⁎ 153.253⁎⁎⁎ 676.954⁎⁎⁎ CFI .914 .903 .886 .970 .939 TLI .920 .947 .866 .981 .965 RMSEA .102 .094 .157 .054 .153 Model Comparison Test χ2 1.228, NS 48.481⁎⁎⁎ 0.273, NS 36.173⁎⁎⁎ 158.054⁎⁎⁎ Note: All factor loadings significant at p ≤ .001. ⁎⁎⁎ p ≤ .001. Table options Results for the other latent constructs are similar. For example, while the model comparison test for the vandalism construct does reach significance (Δ χ2 = 48.481, p < .001) and the two-factor model does fit the data more closely, the latent constructs of perceptual peer vandalism and actor self-reported vandalism are very similar (r = .844). For the violence construct, results closely mirror the theft models; the one and two-factor solutions for violence fit the data equally as well, and the constructs are not significantly different from each other (Δ χ2 = 0.273, p = .60). For the drug-related models, the one-factor model fits the data quite closely, but the two-factor model does provide a significant improvement of fit over the one-factor solution. While the model comparison test is significant (Δ χ2 = 36.173, p < .001), the factors of perceptual peer drug behavior and actor self-reported drug behavior are very similar (r = .877). The final model in Table 3 compares one- and two-factor solutions of perceptual peer alcohol-related behaviors and actor self-reported alcohol-related behaviors. Of particular note, the value of the model comparison test is much larger than the prior models (Δ χ2 = 158.054, p < .001) even though the correlation between the factors of perceptual peer alcohol-related behavior and the actor’s own alcohol-related behavior is large (r = .856). Thus, the second series of CFA models reveals a set of results that provide partial support for hypothesis two – the latent constructs of perceptual peer delinquency and actor self-reported delinquency are very similar to one another across each of the five constructs examined. In the case of the theft and violence constructs, there is no statistical difference between one-factor and two-factor models. It is also noteworthy that the two constructs are very highly correlated across each of the considered behavioral types, which suggests that there is minimal discriminant validity between the actor’s self-reported delinquency and perceived peer delinquency. Comparing actor self-reported delinquency to partner self-reported delinquency The final set of CFA models estimated are shown in Table 4 and compare one- and two-factor solutions between the actor’s and the partner’s self-reported delinquency. Beginning with theft, the factor loadings are all highly significant, but the one-factor model does not show evidence of close fit to the data. However, the two-factor model fits the data so closely that the goodness of fit χ2 fails to reach statistical significance (χ2 = 23.357, p = .08). The correlation between the latent constructs of actor and peer self-reported theft is also well below 1.00 (r = .345). In short, this provides strong evidence that an actor’s self-reported theft is a distinct construct from the peer’s theft behavior, providing initial support for hypothesis three. Table 4. Confirmatory Factor Analysis Models Loading Actor Self-Reported Delinquency (ASRD) and Peer Self-Reported Delinquency (PSRD) Measures onto One and Two Constructs (n = 2,154) Theft Vandalism Violence Drugs Alcohol 1 Factor 2 Factor 1 Factor 2 Factor 1 Factor 2 Factor 1 Factor 2 Factor 1 Factor 2 Factor Item # λ λ λ λ λ λ λ λ λ λ ASRD 1 .712 .855 / - .705 .787 / - .589 .670 / - .763 .864 / - .825 .877 / - ASRD 2 .769 .837 / - .809 .874 / - .606 .748 / - .723 .758 / - .615 .684 / - ASRD 3 .469 .574 / - .785 .902 / - .556 .693 / - .817 .881 / - .580 .653 / - ASRD 4 .490 .591 / - .501 .572 / - .893 .919 / - .864 .907 / - ASRD 5 .431 .537 / - .723 .786 / - PSRD 1 .697 - / .853 .704 - / .785 .600 - / .680 .774 - / .873 .803 - / .873 PSRD 2 .753 - / .832 .811 - / .876 .604 - / .743 .721 - / .755 .611 - / .680 PSRD 3 .469 - / .582 .785 - / .905 .560 - / .692 .820 - / .882 .572 - / .646 PSRD 4 .464 - / .570 .513 - / .583 .874 - / .905 .846 - / .896 PSRD 5 .432 - / .536 .743 - / .813 2 Factor Correlation .345*** .267*** .347*** .590*** .591*** 1 Factor Model Statistics χ2 453.313*** 768.456*** 462.782*** 382.669*** 1122.820*** CFI .752 .773 .773 .637 .889 TLI .682 .735 .735 .533 .889 RMSEA .125 .144 .144 .180 .235 2 Factor Model Statistics χ2 23.357, NS 49.641*** 17.002* 57.751*** 65.663*** CFI .995 .991 .992 .882 .995 TLI .994 .991 .990 .896 .996 RMSEA .017 .027 .027 .107 .045 Model Comparison Test χ2 207.752*** 298.752*** 244.710*** 157.549*** 495.370*** Note: All factor loadings significant at p ≤ .001. * p ≤ .05 *** p ≤ .001. Table options For the vandalism and violence latent constructs, a similar story emerges. Consistently, the two friends’ self-reported delinquencies load on different factors; the two-factor models’ fit indices indicate very close fit to the data, and the highly significant model comparison tests show that a two-factor model is much more appropriate than a one-factor model (Δ χ2 = 298.752 for vandalism; Δ χ2 = 244.710 for violence). Further, the respective constructs of actor and peer self-reported vandalism and violence are moderately correlated with one another (r = .267 for vandalism; r = .347 for violence), but again are well below 1.00. In short, the actor’s and the peer’s self-reported vandalism and violence behaviors are clearly two distinct constructs, providing additional support for hypothesis three. The remaining two latent constructs – drug-related and alcohol-related behaviors – are presented in the final two columns of Table 4. As in the prior models, the two-factor solutions for both constructs provide a significant improvement of fit to the data (Δ χ2 = 157.249 for drugs; Δ χ2 = 495.370 for alcohol). However, the correlations between the actor’s and peer’s self-reported delinquency are larger in these models than in the theft, vandalism and violence models, suggesting that there is a higher degree of similarity between the actor and the partner’s drug and alcohol-related behavior (r = .590 for drugs; r = .591 for alcohol). Overall, the cumulative results in Table 4 reveal consistent support for hypothesis three and indicate that the actor’s and the peer’s self-reported delinquency are two separate constructs. Thus, the use of a measure of peer delinquency obtained directly from peers is an operationally valid measure of peer delinquency that bears strong discriminant validity when compared against respondent delinquency. To determine the degree of support for hypothesis four, we compared the two factor correlations for each of the latent constructs between Table 3 and Table 4. These comparisons provide strong support for hypothesis four and are consistent with prior research (Haynie and Osgood, 2005, Rebellon and Modecki, in press and Young et al., in press): The strength of the relationship between actor self-reported delinquency and partner delinquency is substantively smaller in magnitude when the measure comes directly from the partner. This difference in magnitude is most evident for the vandalism construct. Specifically, the two factor correlation between actor and partner self-reports is 0.267, while the correlation between actor self-reported vandalism and the actor's perception of partner vandalism is 0.844, which represents over a threefold difference in magnitude. Similar, but less dramatic, differences are also observed for each of the four remaining delinquency constructs. Thus, in these data of friendship dyads, the strength of the relationship between actor and partner delinquency is particularly sensitive to the manner in which peer delinquency is operationalized.